FeatherRouter

An intelligent multi-agent router that breaks down coding tasks and dynamically assigns the best open-source model from Featherless API for every build stage.

Comment

Inspiration

Most AI coding assistants rely on a single monolithic LLM for every stage of development. However, empirically, no single model excels at everything — vision-capable models (like Qwen3-VL) excel at architecture layout from wireframes, specialized code models (like Qwen2.5-Coder-32B) excel at syntax synthesis, and instruct models (like Mistral-Small) excel at security auditing.

Forces a single model to handle all stages degrades output quality and increases error rates. We built FeatherRouter to create a dynamic multi-agent model router that breaks software generation into specialized pipeline stages and assigns the optimal open-source model per stage across 30,000+ models available on the Featherless API.

What It Does

FeatherRouter is an intelligent model orchestration platform for web application generation:

  1. Dynamic Task Analysis: Evaluates incoming user prompts and wireframe images using Gemini 2.5 Flash as a cognitive router.
  2. Stage-Aware Model Selection: Scores and ranks candidate models from the Featherless inventory based on capabilities, context window, and latency.
  3. Multi-Stage Execution: Dispatches specialized tasks across pipeline stages (Architecture Plan → Code Synthesis → Security Audit).
  4. Automated Quality Review (Layer 4): Audits cross-file DOM element matching, CSS contrast, and JS event listener bindings before rendering.
  5. Live Workspace & Web Preview: Provides an interactive code editor, real-time iframe browser preview with DOM polyfills, AI refinement bar, and 1-click ZIP export.

How We Built It

  • Frontend & Canvas: Built with Next.js 16 (App Router), React 18, and TypeScript. Hand-crafted dark mode styling with zero external UI fluff.
  • Cognitive Router Brain: Gemini 2.5 Flash (via Google AI Studio) acts as the high-speed routing head and automated quality reviewer.
  • Model Execution Engine: Integrated with the Featherless API, accessing over 21,700+ open-source models (including Qwen2.5-Coder-32B, DeepSeek-R1, and Mistral-Small).
  • Client-Side Preview Polyfill: Built a custom iframe bundler injecting DOMContentLoaded execution polyfills, smooth anchor scrolling, and link navigation interceptors.

Mathematical Scoring Model & Technical Challenges

To select candidate model $M_i$ for pipeline stage $T_k$, FeatherRouter evaluates a weighted fitness score:

$$S(M_i, T_k) = w_1 \cdot C(M_i) + w_2 \cdot L(M_i) + w_3 \cdot P(M_i, T_k)$$

Where $C(M_i)$ is context window capacity, $L(M_i)$ is observed inference latency, and $P(M_i, T_k)$ is historical task performance alignment.

Key Technical Challenges Solved:

  • Client-Side Iframe Navigation Leaks: Solved link navigation bugs where clicking <a> tags inside generated previews reloaded the parent application by implementing event capture interceptors with smooth anchor scrolling.
  • Model Output Deduplication: Resolved LLM code block repetition by detecting top-level entry point triggers (DOMContentLoaded, initGame).
  • File Extraction Fallbacks: Built regex parsers that auto-extract embedded <style> and <script> tags into distinct styles.css and script.js files if an LLM returns a single monolithic HTML file.

Accomplishments That We're Proud Of

  • Shipped a 100% functional live production app deployed on Firebase Hosting with zero TypeScript compilation errors.
  • Built a fully transparent routing drawer displaying real-time model marks ($/100$), candidate rankings, and natural-language selection rationale.
  • Successfully orchestrated 4 distinct model stages with automatic fallback queues.

What We Learned

  • Open-source models (like Qwen2.5-Coder-32B) rival proprietary models when given specialized, scoped task prompts rather than general instructions.
  • Cross-file state alignment (matching JS document.getElementById with HTML id="...") requires an explicit automated QA pass.

What's Next for FeatherRouter

  • Custom Benchmark Suite: Expanding automated evaluation metrics for real-time model scoring.
  • Multi-Framework Exports: Adding native support for Vue, Svelte, and React Native codebases.
  • Local Model Routing: Supporting local Ollama / LM Studio endpoints alongside Featherless API cloud inference.

Built With

Share this project:

Updates

Submission history